Atinamos Evidence & Findings

This page is an index of public observations. It intentionally separates strong empirical findings from development-only evidence and open questions.

You can pay to be tested. You cannot pay to be trusted.

Specific named-service verification cases now live in the dedicated Verification Studies section. This page consolidates what those studies and earlier experiments currently support.

Independent external verification studies — 22 August 2026

Atinamos exercised three independently owned machine services through real x402 payment paths using a generic recipe-driven verifier. The three observations produced three different outcomes.

Study 1 — x402Node JSON Repair

Observation: The verifier observed a live x402 v2 Base payment requirement, paid 0.006 USDC, received a synchronous JSON result and applied four objective assertions. All four passed.

Supports: an independently owned paid service can be exercised and validated through the generic Atinamos recipe runner.

Does not support: permanent trust, repeated reliability, autonomous discovery or proof that the same recipe model works for every service category.

Read the full x402Node study.

Study 2 — x402.direct Service Directory Search

Observation: A live x402 v1 challenge advertised 0.001 USDC on Base but contained trailing whitespace in payTo. Atinamos preserved the anomaly, safely normalised surrounding whitespace and created the required legacy payment authorization. The authorised request returned HTTP 500. A read-only Base USDC check found no 0.001 USDC settlement.

Supports: machine-contract defects and paid-path failures can be detected before a buyer incorrectly treats a service as successfully fulfilled.

Does not support: a claim that x402.direct always fails, or that the observed HTTP 500 was caused by the whitespace anomaly.

Read the full x402.direct study.

Study 3 — x402engine Web Screenshot

Observation: The verifier paid 0.01 USDC on Base and independently confirmed settlement on-chain. The service returned HTTP 200 and page metadata consistent with the target page. However, the primary screenshot field did not decode as the advertised base64 PNG, so the verifier recorded output_valid=false.

Supports: payment settlement and HTTP 200 fulfilment are not sufficient evidence that the seller delivered the advertised output contract.

Does not support: a claim that x402engine always produces an invalid screenshot or that the provider as a whole is unsafe or untrustworthy.

Read the full x402engine study.

Machine-readable buyer evidence

The same published evidence is available to software through:

GET /v1/trust?endpoint=<service-url>

The lookup returns timestamped observations, settlement/fulfilment state, validation outcome, classification and, where available, a link to the public Study page. It deliberately does not return a universal trust verdict. A buyer agent decides whether to proceed, reject, request a fresh test or apply additional constraints according to its own policy.

Earlier publishable findings

Virtuals ACP — paid service execution

Observation: A real Atinamos Render Check job completed through Virtuals Agent Commerce Protocol and escrow was released.

Supports: Atinamos can operate as a paid service provider in ACP.

Does not support: an independent buyer autonomously discovered and selected Atinamos without provider preselection.

x402 / Coinbase Bazaar — payable and discoverable seller path

Observation: Atinamos Render Check has operated behind an x402 payment requirement and has been visible through Coinbase/x402 discovery data.

Supports: Atinamos has a real machine-payable HTTP seller path and can participate in discovery.

Does not support: discovery by itself proving seller quality or autonomous buyer intent.

Circle — external paid purchase for 0.015 USDC

Observation: A policy-constrained Circle agent wallet purchased an external Notte service for 0.015 USDC and received a machine-readable result.

Supports: AI-controlled delegated spending can result in a real external machine purchase.

Limitation: this experiment is evidence about buyer/payment capability, not Atinamos seller trust.

Circle + external evidence — x402Node JSON Repair for 0.006 USDC

Observation: ResolveBots evidence identified a candidate; a current executable contract check established the required POST route; the buyer was capped at 0.01 USDC; 0.006 USDC was paid on Base; repaired JSON was returned and validated.

Supports: historical evidence, live contract inspection, spend policy, payment and output validation can form one procurement decision sequence.

Limitation: that earlier orchestration was human-guided. The later Atinamos verifier study independently exercised the same external service and is documented separately in Studies.

Society AI — task execution without demonstrated settlement

Observation: Society AI successfully routed a real task into Atinamos Render Check and received structured results. The tested calls did not reduce the buyer balance, creator earnings/direct payments were not active in the observed product state, and logged requests were human-chat initiated with no delegating agent.

Supports: agent-network messaging and self-hosted remote execution worked.

Does not support: paid autonomous agent procurement.

Development evidence — not a public service-quality claim

Atinamos Verification contract fingerprinting

The verifier can register an exact endpoint, call it, parse its x402 payment requirement, fingerprint contract fields and compare later observations. During development the endpoint was deliberately changed and the fingerprint changed; restoring it returned the original fingerprint.

This validates the prototype's change-detection mechanism. It should not be counted as independent service-quality evidence.

Next public evidence target — Decision Delta

The next decisive question is no longer whether Atinamos can pay and validate an external service. It can.

The next test is whether independent evidence changes procurement behaviour:

candidate services → Atinamos evidence → buyer policy → choose/reject → payment

A useful Decision Delta occurs when evidence causes a buyer to avoid a costly mistake, justify a premium for stronger evidence or constrain execution/spending conditions.